CN104866831A - Feature weighted face identification algorithm - Google Patents

Abstract

The invention relates to a feature weighted face identification algorithm, belongs to the field of human face technology. The feature weighted face identification algorithm comprises the steps that a face image is decomposed into high and low frequency components via wavelet transformation, principal component analysis (PCA) is carried out on the different components to extract feature images, weighing is carried out by an AHP algorithm according to the significance of the components, and a support vector machine (SVM) is used for classification and identification. The image identification rate is highest, and all main information of images is used for identification.

Description

The face recognition algorithms of characteristic weighing

Technical field

The present invention relates to face technical field, be specifically related to a kind of face recognition algorithms of characteristic weighing.

Background technology

The application of current face's recognition technology is more and more extensive, as: face recognition door control system, supervisory system etc.It has become a focus received much concern [1] [2] in artificial intelligence and pattern identification research field.But face recognition algorithms still has the place of many improvement, as feature extraction, dimension control, recognition accuracy etc.

Because the dimensional comparison of facial image is high, common way facial image is carried out dimensionality reduction extract eigenface, then compare.Wherein principal component analysis (PCA) (PCA) is exactly carry out dimension-reduction treatment to image, obtain the major component of facial image, remove the correlativity of raw data, generating feature face, test pattern and eigenface are compared again and identify, the method has achieved good recognition effect [3].But application PCA method, easily ignores other ingredients of image, identifies that accuracy still has much room for improvement.After this people also propose to use sorter, sort out face, classifying to human face data as support vector machine (SVM) and return, the method usable range is wide, can process any data, but the face characteristic value of its accuracy and input has comparatively Important Relations.Based on the face identification method of wavelet transformation, by the multi-resolution decomposition of image, face can be resolved into low-and high-frequency different piece, usually get low frequency part that image information enriches to carry out recognition of face comparison the method and achieve certain effect, improve recognition accuracy, but eliminate the HFS of image, cause the loss of partial information.

Summary of the invention

Solve the problems of the technologies described above, in order to more fully extract the feature of facial image, the present invention proposes a kind of face recognition algorithms of characteristic weighing, first adopt wavelet transformation that facial image is resolved into high-low frequency weight, then principal component analysis (PCA) (PCA) is carried out to different component and extract characteristic image, use AHP algorithm to be weighted according to the importance of each component again, finally use support vector machine (SVM) to carry out Classification and Identification.The method carries out simulating, verifying on the face database of classics, and recognition effect is obviously better than traditional recognition methods.

In order to achieve the above object, the technical solution adopted in the present invention is, the face recognition algorithms of characteristic weighing, comprises the following steps:

Adopt wavelet transformation that facial image is resolved into high-low frequency weight,

Principal component analysis (PCA) (PCA) is carried out to different component and extracts characteristic image,

AHP algorithm is used to be weighted according to the importance of each component,

Support vector machine (SVM) is used to carry out Classification and Identification.

Further, adopt wavelet transformation that facial image is resolved into high-low frequency weight, specifically comprise:

If ψ (t) ∈ is L ²(R) be a quadractically integrable function, if ψ (t) meets following condition:

$\underset{R}{\&Integral;}\frac{\left|\mathrm{\&psi;}{\left(\mathrm{\&omega;}\right)}^2\right|}{\mathrm{\&omega;}}\mathrm{d\&omega;}<\&infin;---\left(1\right)$

Then claim ψ (t) to be a wavelet function, and claim formula (1) to be the admissible condition of wavelet function, flexible and translation wavelet mother function ψ (t), obtains wavelet basis function:

${\mathrm{\&psi;}}_{a,\mathrm{\&tau;}}\left(t\right)=a^{-1/2}\mathrm{\&psi;}\left(\frac{t-\mathrm{\&tau;}}{a}\right)---\left(2\right)$

Wherein, a and τ is real number, and a>0, a are contraction-expansion factor, and τ is shift factor,

Function f (t) ∈ L ²(R) continuous wavelet transform CWT is defined as follows:

$W{T}_f(a,\mathrm{\&tau;})=\frac{1}{\sqrt{a}}\underset{R}{\&Integral;}f\left(t\right)\stackrel{\&OverBar;}{\mathrm{\&psi;}\left(\frac{t-\mathrm{\&tau;}}{a}\right)}\mathrm{dt}---\left(3\right)$

In formula for the conjugate function of mother wavelet, can obtain a series of wavelet coefficient by formula (3), these coefficients are functions of shift factor and zoom factor.

Further, wavelet transform is expressed as:

$a_0^{-m/2}\underset{R}{\&Integral;}f\left(x\right)\mathrm{\&psi;}\left(\frac{x-n{b}_0}{a_0^m}\right)\mathrm{dx}=\&lang;f,{\mathrm{\&psi;}}_{m,n}\&rang;---\left(4\right)$

Wherein, a ₀, b ₀for constant, and a ₀>0, m, n are integer.

Further, principal component analysis (PCA) (PCA) is carried out to different component and extracts characteristic image, specifically comprise:

If the size of facial image is m × n, after vectorization, become the column vector of M=m × n dimension, face training sample is N, X _ibe the column vector of i-th sample, then get training sample average value mu:

$\mathrm{\&mu;}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{X}_i---\left(5\right)$

After more each training sample being deducted face average, composition matrix A=[X ₁-μ, X ₂-μ ..., X _n-μ], so the covariance matrix of training sample is:

$C=\frac{A\&CenterDot;A^T}{N}---\left(6\right)$

The proper vector asked for again corresponding to C nonzero eigenvalue forms the optimum projection subspace that will find, and in the recognition of face of reality, general eigenwert must accumulation contribution rate determine the major component dimension d that will choose, generally choose the eigenwert characteristic of correspondence vector structural attitude space making α>=90%, then the matrix of feature space is U=[u ₁, u ₂..., u _d], training sample is projected on feature space, obtains projection matrix:

Q＝U ^TA (7)

Be the eigenface of sample.

Further, use AHP algorithm to be weighted according to the importance of each component, specifically comprise:

1., first problem stratification, construct a stratified structural model;

2. Judgement Matricies, the relation between hierarchical structure reflection factor, but each criterion proportion in target is weighed is different;

3. Mode of Level Simple Sequence and consistency check, determines the associated element importance of this level, and the weighted value between their order.

Further, support vector machine (SVM) is used to carry out Classification and Identification, specifically comprise: select suitable support vector machine parameter, extract face characteristic data label, the face characteristic data extracted are carried out training and obtains training set, again test sample book is supplied to support vector machine, provides recognition result by the supporting vector machine model trained.

The present invention is by adopting technique scheme, and compared with prior art, tool has the following advantages:

The face recognition algorithms of characteristic weighing is proposed herein, first adopt wavelet transformation that facial image is resolved into high-low frequency weight, then principal component analysis (PCA) (PCA) is carried out to different component and extract characteristic image, use AHP algorithm to be weighted according to the importance of each component again, finally use support vector machine (SVM) to carry out Classification and Identification.Image recognition rate of the present invention is the highest, and the main information of image is all used to identify.

Accompanying drawing explanation

Fig. 1 is the chain graph of the system of the embodiment of the present invention.

Fig. 2 is the topological diagram of the system of the embodiment of the present invention.

Fig. 3 is each component of one deck wavelet decomposition and subimage thereof.

Fig. 4 is the basic thought schematic diagram of support vector machine.

Fig. 5 is algorithm flow chart.

Fig. 6 is face experimental data base figure

The Comparative result figure that Fig. 7 (I) algorithms of different compares.

The Comparative result figure that Fig. 7 (II) algorithms of different compares.

Fig. 8 feature accumulated value affects schematic diagram to discrimination

Embodiment

Now the present invention is further described with embodiment by reference to the accompanying drawings.

As a specific embodiment, as depicted in figs. 1 and 2, a kind of novel examination face authentication system of the present invention, comprise data collector, data processing equipment, cloud storage server and multiple admission checking harvester, data collector is set up data communication by wired or wireless internet and data processing equipment and is connected, data processing equipment is set up data communication by wired or wireless LAN (Local Area Network) with cloud storage server and is connected

Data collector gathers character information and image information and generates image data, is sent to data processing equipment in real time; Data processing equipment according to the region affiliation of image data ground, by the territorial classification database of correspondence in acquired data storage to the remote port database of cloud storage server;

Multiple admission checking harvester is distributed in regional ownership place admission check post, checking harvester in field gathers Data Concurrent to be tested and delivers to cloud storage server, cloud storage server is adjusted and these data to be tested and image data is compared, whether mate with the territorial classification database at corresponding image data place with judging the region affiliation of these data to be tested, matching result generates authorization information, to judge that can these data to be tested obtain verifying authorization, if, then further by 1 than the relative method of N to knowing whether data to be tested mate with image data, and then obtain these data to be tested whether by the admission checking on described region affiliation ground.

Another technical scheme of the present invention is, a kind of novel examination face authentication method, comprises the following steps:

Data collector gathers character information and image information and generates image data, is sent to data processing equipment in real time; Data processing equipment according to the region affiliation of image data ground, by the territorial classification database of correspondence in acquired data storage to the remote port database of cloud storage server;

Multiple admission checking harvester is distributed in regional ownership place admission check post, checking harvester in field gathers Data Concurrent to be tested and delivers to cloud storage server, cloud storage server is adjusted and these data to be tested and image data is compared, whether mate with the territorial classification database at corresponding image data place with judging the region affiliation of these data to be tested, matching result generates authorization information, to judge that can these data to be tested obtain verifying authorization, if, then further by 1 than the relative method of N to knowing whether data to be tested mate with image data, and then obtain these data to be tested whether by the admission checking on described region affiliation ground.

As shown in Figure 1, be the chain graph of the system of the embodiment of the present invention, its adopts the mode directly uploading, access, identify from bottom to top, namely user terminal can directly and the webserver carry out real-time, interactive.This system data acquisition device is integrated in PC, flat board, mobile phone, can realize the collection of face, storage, identification by data collector; Data collector adopts the modes such as Mobile phone card flow, wireless routing, broadband network to transmit, data processing equipment is then web-transporting device, this web-transporting device has fire wall, load-equalizing switch (realizing multiple terminal access not time delay) etc., Large Volume Data is effectively transmitted, prevents from distorting, attacking.Cloud storage server comprises two-stage, and one-level is province, site, city server, effectively stores, accesses, issues data; Another Ji Wei great district webserver, carries out large data analysis to the data that each department are uploaded, and statistics examinee information, issues comprehensive examination district information in time.

This system allows user to carry out personalized customization, provides document content editor, edit model, reaches the effect of What You See Is What You Get.This system should have good security, extensibility, supports more high traffic by hardware or software upgrading.System adopts the method for designing of modularization, modularization (namely user can change recognizer), objectification, easy of integration, easily customize, and possesses good ability of second development, really makes the investment of user minimum, the Maximum Value of creation.System provides daily management maintenance, easily extensible, real-time, time delay is little, recognition accuracy is high.

This system meets the specification of national test office, is applicable to dissimilar invigilator's certification, accurately identifies face information, and high speed storing retrieval face resource, guarantees Information Security, provide the use habit of different crowd, network environment etc.

As shown in Figure 2, be the topological diagram of this system, data center stores the human face data information in urban district, each province, and data center of provinces and cities carries local face data message, and handheld terminal is direct this data message of comparison just.

The feature of this system is the comparative approach adopting 1:N, and wherein N represents different N and opens face, and 1 representative needs the face of comparison.Traditional authentication method, is the method adopting 1:1, namely the examination data of server end is downloaded to local terminal, then the examination of the picture downloaded and admission is contrasted, judges whether examinee according to comparison result.Such as during examinee A admission, local terminal has downloaded the information of examinee A in advance, when examinee A admission, and the photo that comparison is downloaded, and I.There is significant deficiency in the method because download to the picture of local terminal from server, can be artificial distort picture, the mistake of authentication information will be caused like this, may exist and impersonate phenomenon.And by the way that this system proposes, effectively can evade so a kind of phenomenon, because N opens different picture-storage at server end, local end user, without the data of weight update server end, when comparison, examinee needs the N of comparison simultaneously to open different picture contrasts.Such as, after examinee A enters examination hall, handheld terminal logon server website, comparison examinee A, if server end can find out the information of examinee A, then can admission, if do not existed, then not me.Examination face authentication method in the past, adopt 1:1 comparative approach, namely the face picture of test taker number there is is to download to local terminal server end, oneself test taker number is shown in now examinee's admission, the human face photo having downloaded concrete test taker number can be navigated to, and then comparison examinee and corresponding photo, the method location is fast, comparison speed is fast, is usually used in examination admission certification.But the drawback of the method also obviously, namely invigilator person may revise the local terminal photo downloaded, and photo will be caused like this to change, thus test-taker is had an opportunity to take advantage of, and therefore adopts the method to have more potential safety hazard.The 1:N method that this embodiment proposes, can local information be effectively avoided to be tampered, namely during examinee's admission, invigilator person gathers examinee's photo, and directly connects backstage face database, compare, open different examinee's face information because backstage face database has N, be therefore the comparison of 1 couple of N, the method effectively prevent examinee information and is tampered, avoid examinee information download after comparison again, reach that foreground gathers, the effect of backstage high speed comparison.

This system also has an advantage, transregionally exactly set up different comparison data storehouses, and namely the Verification System in somewhere can only log in this area.Certain examination hall, section of such as Licheng District, Quanzhou, can only the corresponding database logging in this section, can effectively prevent database access amount large like this, comparison problem not in time.And the central server of Licheng District can check the examination hall authentication scenario under all sections, this district, in like manner the examination authentication scenario in all districts can be checked by Quanzhou City.The backstage comparison that the present embodiment proposes, compare in partition territory, namely which section the examinee gathered belongs to, with regard to examinee information comparison that is direct and this section, prevent the face information comparison of large data, such as: examination district, Licheng District first, examinee A admission Quanzhou, the hand-held dull and stereotyped instrument of invigilator person gathers examinee A, now examinee A only with whole examinee information comparisons of this section, and do not compare with citywide examinee information, avoid large-scale comparison repeatedly, and the reduced time is shorter, admission of taking an examination can be applicable to.

In the present embodiment, described 1 is specially than the relative method of N: 1 data to be tested and N number of image data are compared by the face recognition algorithms of characteristic weighing by cloud storage server, the face recognition algorithms of this characteristic weighing comprises the following steps: first adopt wavelet transformation that facial image is resolved into high-low frequency weight, then principal component analysis (PCA) (PCA) is carried out to different component and extract characteristic image, use AHP algorithm to be weighted according to the importance of each component again, finally use support vector machine (SVM) to carry out Classification and Identification.

Wavelet transformation

Wavelet transformation is developing rapidly nearly ten years, and it is extended out by Fourier transform, can provide multiresolution and multiscale analysis, and it obtains application of result [4] in image processing and analyzing, computer vision, signal transacting etc.

Wavelet transformation is proposed, if ψ (t) ∈ is L in first time in 1984 by researchers such as Morlet ²(R) be a quadractically integrable function, if ψ (t) meets following condition:

$\underset{R}{\&Integral;}\frac{\left|\mathrm{\&psi;}{\left(\mathrm{\&omega;}\right)}^2\right|}{\mathrm{\&omega;}}\mathrm{d\&omega;}<\&infin;---\left(1\right)$

Then claim ψ (t) to be a wavelet function, and claim formula (1) to be the admissible condition of wavelet function, flexible and translation wavelet mother function ψ (t), obtains wavelet basis function:

${\mathrm{\&psi;}}_{a,\mathrm{\&tau;}}\left(t\right)=a^{-1/2}\mathrm{\&psi;}\left(\frac{t-\mathrm{\&tau;}}{a}\right)---\left(2\right)$

Wherein, a and τ is real number, and a>0, a are contraction-expansion factor, and τ is shift factor.

Function f (t) ∈ L ²(R) continuous wavelet transform CWT is defined as follows:

$W{T}_f(a,\mathrm{\&tau;})=\frac{1}{\sqrt{a}}\underset{R}{\&Integral;}f\left(t\right)\stackrel{\&OverBar;}{\mathrm{\&psi;}\left(\frac{t-\mathrm{\&tau;}}{a}\right)}\mathrm{dt}---\left(3\right)$

In formula for the conjugate function of mother wavelet.Can obtain a series of wavelet coefficient by formula (3), these coefficients are functions of shift factor and zoom factor.

In practical application, when processing small echo, usually it is desirable that discrete signal, now just need the size changing factor a and continuous translation parameter τ, the analysis of signal on different scale can not only be met like this, selecting scale can also be carried out according to different objects.This analytical approach is very effective, and result is also very accurate.Wavelet transform can be expressed as:

$a_0^{-m/2}\underset{R}{\&Integral;}f\left(x\right)\mathrm{\&psi;}\left(\frac{x-n{b}_0}{a_0^m}\right)\mathrm{dx}=\&lang;f,{\mathrm{\&psi;}}_{m,n}\&rang;---\left(4\right)$

Wherein, a ₀, b ₀for constant, and a ₀>0, m, n are integer.

The method taked herein carries out one deck wavelet transform to facial image, generates horizontal component, vertical component and diagonal components, and as shown in Figure 3, what obtain is 4 subgraph components of face.

In Fig. 3, LL is low frequency component, contains most information of former figure, and LH is the confidence of people face, and HL is vertical component, and comprise the marginal informations such as the nose of people, ear, HH is diagonal components, comprises information less.

2.2 principal component analysis methods

Principal component analysis (PCA) (PCA) is a kind of conventional Mathematical Method, it is the sample point certain correlativity, choose the maximum direction of these sample point variances as feature space, reconstitute one group of incoherent data, thus packed data [3].

If the size of facial image is m × n, after vectorization, become the column vector of M=m × n dimension, face training sample is N, X _ibe the column vector of i-th sample, then get training sample average value mu:

$\mathrm{\&mu;}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{X}_i---\left(5\right)$

After more each training sample being deducted face average, composition matrix A=[X ₁-μ, X ₂-μ ..., X _n-μ], so the covariance matrix of training sample is:

$C=\frac{A\&CenterDot;A^T}{N}---\left(6\right)$

The proper vector asked for again corresponding to C nonzero eigenvalue forms the optimum projection subspace that will find, and in the recognition of face of reality, general eigenwert must accumulation contribution rate determine the major component dimension d that will choose, generally choose the eigenwert characteristic of correspondence vector structural attitude space making α>=90%.Then the matrix of feature space is U=[u ₁, u ₂..., u _d], training sample is projected on feature space, obtains projection matrix:

Q＝U ^TA (7)

Be the eigenface of sample.

2.3AHP algorithm

Analytical hierarchy process (AHP) is that the U.S. plans strategies for scholar T.L.Saaty professor in the proposition of phase early 1970s, and AHP provides criteria decision-making method to a challenge.It needs the structural model setting up Recurison order hierarchy, Judgement Matricies, then carries out Mode of Level Simple Sequence and consistency check [5], and the performing step of this algorithm is as follows:

1., first problem stratification, construct a stratified structural model;

2. Judgement Matricies, the relation between hierarchical structure reflection factor, but each criterion proportion in target is weighed is different;

3. Mode of Level Simple Sequence and consistency check, determines the associated element importance of this level, and the weighted value between their order.

The key of AHP algorithm is Judgement Matricies, and it is according to the relation between factor, and give different weight, reference numerals 1-9 and inverse thereof are as scale here, and table 1 lists the implication of 1-9 scale.

The scoring criteria of element in table 1 judgment matrix

Judging between two by making n (n-1)/2 time to element each in matrix, can judgment matrix be derived.

$A=\left[\begin{array}{ccccc}1& a_{12}& .& .& a_{1n}\\ a_{21}& 1& .& .& a_{2n}\\ a_{31}& a_{32}& 1& .& a_{3n}\\ .& .& .& .& .\\ a_{n1}& a_{n2}& .& .& 1\end{array}\right]---\left(8\right)$

Again corresponding Maximum characteristic root λ is asked for judgment matrix A _max:

Aω＝λ _maxω (9)

The component of ω is exactly the weighted value of the single sequence of corresponding factor.

Need in addition to carry out consistency check to weighted value, order matrix B=(b _ij) _{n × n}, wherein (i, j=1,2 ... n)

Make again obtain judgment matrix ordering vector w thus _i=(w ₁, w ₂... w _n) ^tmethod be called " Sum-Product algorithm ".Now the eigenvalue of maximum of judgment matrix A can be approximated to be calculate coincident indicator again:

$\mathrm{CI}=\frac{{\mathrm{\&lambda;}}_{\max }-n}{n-1}---\left(10\right)$

Finally calculate consistency ration wherein the value of RI is in table 2.

Table 2 RI span

n	1	2	3	4	5	6	7	8	9	10	11
												RI	0	0	0.58	0.90	1.12	1.24	1.32	1.41	1.45	1.49	1.52

Just think that as CR<0.1 total hierarchial sorting result has comparatively satisfied consistance and accepts this analysis result, otherwise just do not accept.Here you need to add is that, after consistency check completes, also need, to judgment matrix inspection ordinal consistency, finally to reach crash consistency.

2.4 support vector machine

Support vector machine grows up on the basis of Statistical Learning Theory, it is a learning model having supervision, be commonly used to carry out pattern-recognition, classification and regretional analysis, its main thought is that the input space is mapped to a higher dimensional space by the nonlinear transformation defined by interior Product function, the data of original linearly inseparable are made to become the data [6] of linear separability, and then solve the optimal separating hyper plane of higher dimensional space, as shown in Figure 4.

Select suitable support vector machine parameter, extract face characteristic data label, the face characteristic data extracted are carried out training and obtains training set, then test sample book is supplied to support vector machine, provide recognition result by the supporting vector machine model trained.

3 algorithm realization

Utilized wavelet transformation and principal component analysis (PCA) mainly to extract low frequency component in the past and removed high fdrequency component, and on low frequency component, directly applied PCA algorithm extract face characteristic, then carry out svm classifier identification.The weak point of the method is, directly remove the high-frequency information of face picture, identification division can be made imperfect, and every part of image all plays a role to identification, make full use of the useful information of each several part.

In view of above-mentioned analysis, method in this paper takes into full account facial image different piece, and implementation step is as follows:

1. first facial image is become low-and high-frequency 4 components through one deck wavelet decomposition;

2. then principal component analysis (PCA) (PCA) is carried out to different component and extract characteristic image;

3. use AHP algorithm to be weighted fusion according to the importance of each component again;

4. finally using the image after merging as face characteristic, more all samples are divided into training set and test set, use SVM to carry out Classification and Identification, its realization flow as shown in Figure 5:

The weight calculation formula of this algorithm is as follows:

X＝ω ₁LL+ω ₂LH+ω ₃HL+ω ₄HH (11)

Here 4 weights are the importance according to different component, calculate, wherein ω according to AHP algorithm ₁+ ω ₂+ ω ₃+ ω ₄=1.

4 experimental results and analysis

This algorithm is tested at classical face database, and this database has AT & T, ORL, Yale etc.As shown in Figure 6, be the part face information of this database.

In experiment, first the gray level image of face database is carried out pre-service, picture format unification is 112 × 92, and the face details of often opening figure all has difference.Then according to algorithm in this paper, wavelet transformation is carried out to image, generate four width PCA subgraphs, according to AHP algorithm synthesis subgraph, then carry out svm classifier identification.

Experiment 1: in order to verify algorithm in this paper, calculate recognition accuracy under different weights, research weights are on the impact of discrimination.Each class number of training N=5 is selected, α=90% in experiment, then the weights calculating different subgraph according to AHP algorithm, carry out the experiment of many group weights, experimental result is as follows:

Discrimination under the different weights of table 3 this paper algorithm

As can be seen here, as the weights ω of low frequency part ₁during increase, when the weights of HFS reduce, its recognition accuracy all can increase.Therefore by arranging image different piece weights, the accuracy improving recognition of face is contributed to.

Experiment 2 application algorithm herein, contrast major component analyses the discrimination of method, verifies the accuracy of algorithm herein.In each experiment, the corresponding weighted value of algorithm herein, but select different training sample number N, often kind of algorithms selection 5 groups carries out respectively, all gets α=90%, as table 4 shows at every turn:

The comparing result of table 4 algorithms of different

(I)

(II)

Table 4 (I) is respectively 0.74 in weight, 0.14,0.1,0.02, calculate the discrimination that algorithms of different provides, (II) be respectively 0.65 in weight, 0.15,0.17,0.03, Fig. 7 (I), (II) are corresponding trend map respectively, calculate the discrimination that algorithms of different provides.As can be seen from result above, training sample is more, and accuracy rate is higher, and algorithm in this paper is higher than other two kinds of algorithm accuracys, considers to differentiation the different piece of facial image herein.

The effect of accumulation contribution rate α in PCA eigenwert is compared in experiment 3, the value of α can affect the discrimination of algorithms of different, Experimental comparison PCA+SVM, 2DPCA+SNM, herein algorithm (often kind of Algorithm for Training sample N=5, algorithm weights get 0.74 herein, 0.14,0.1,0.02), concrete outcome is as shown in Figure 8:

As can be seen from Figure 8, contribution rate α can affect the discrimination of algorithm, when α=95%, image recognition rate is the highest, and the main information of image is all used to identify, discrimination all can be caused to reduce when α gets other values, therefore, when carrying out PCA dimensionality reduction, the dimension of main one-tenth composition is selected also to be crucial.

Although specifically show in conjunction with preferred embodiment and describe the present invention; but those skilled in the art should be understood that; not departing from the spirit and scope of the present invention that appended claims limits; can make a variety of changes the present invention in the form and details, be protection scope of the present invention.

Claims (6)

1. the face recognition algorithms of characteristic weighing, is characterized in that: comprise the following steps:

Adopt wavelet transformation that facial image is resolved into high-low frequency weight,

Principal component analysis (PCA) (PCA) is carried out to different component and extracts characteristic image,

AHP algorithm is used to be weighted according to the importance of each component,

Support vector machine (SVM) is used to carry out Classification and Identification.

2. the face recognition algorithms of characteristic weighing according to claim 1, is characterized in that: adopt wavelet transformation that facial image is resolved into high-low frequency weight, specifically comprise:

If ψ (t) ∈ is L ²(R) be a quadractically integrable function, if ψ (t) meets following condition:

$\underset{R}{\&Integral;}\frac{\left|\mathrm{\&psi;}{\left(\mathrm{\&omega;}\right)}^2\right|}{\mathrm{\&omega;}}\mathrm{d\&omega;}<\&infin;---\left(1\right)$

Then claim ψ (t) to be a wavelet function, and claim formula (1) to be the admissible condition of wavelet function, flexible and translation wavelet mother function ψ (t), obtains wavelet basis function:

${\mathrm{\&psi;}}_{a,\mathrm{\&tau;}}\left(t\right)=a^{-1/2}\mathrm{\&psi;}\left(\frac{t-\mathrm{\&tau;}}{a}\right)---\left(2\right)$

Wherein, a and τ is real number, and a>0, a are contraction-expansion factor, and τ is shift factor,

Function f (t) ∈ L ²(R) continuous wavelet transform CWT is defined as follows:

${\mathrm{WT}}_f(a,\mathrm{\&tau;})=\frac{1}{\sqrt{a}}\underset{R}{\&Integral;}f\left(t\right)\stackrel{\&OverBar;}{\mathrm{\&psi;}\left(\frac{t-\mathrm{\&tau;}}{a}\right)}\mathrm{dt}---\left(3\right)$

In formula for the conjugate function of mother wavelet, can obtain a series of wavelet coefficient by formula (3), these coefficients are functions of shift factor and zoom factor.

3. the face recognition algorithms of characteristic weighing according to claim 2, is characterized in that: a series of wavelet coefficient comprises discrete wavelet transform coefficients, and wavelet transform is expressed as:

$a_0^{-m/2}\underset{R}{\&Integral;}f\left(x\right)\mathrm{\&psi;}\left(\frac{x-n{b}_0}{a_0^m}\right)\mathrm{dx}=<f,{\mathrm{\&psi;}}_{m,n}>---\left(4\right)$

Wherein, a ₀, b ₀for constant, and a ₀>0, m, n are integer.

4. the face recognition algorithms of characteristic weighing according to claim 1, is characterized in that: carry out principal component analysis (PCA) (PCA) to different component and extract characteristic image, specifically comprise:

If the size of facial image is m × n, after vectorization, become the column vector of M=m × n dimension, face training sample is N, X _ibe the column vector of i-th sample, then get training sample average value mu:

$\mathrm{\&mu;}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{X}_i---\left(5\right)$

After more each training sample being deducted face average, composition matrix A=[X ₁-μ, X ₂-μ ..., X _n-μ], so the covariance matrix of training sample is:

$C=\frac{A\&CenterDot;A^T}{N}---\left(6\right)$

The proper vector asked for again corresponding to C nonzero eigenvalue forms the optimum projection subspace that will find, and in the recognition of face of reality, general eigenwert must accumulation contribution rate determine the major component dimension d that will choose, generally choose the eigenwert characteristic of correspondence vector structural attitude space making α>=90%, then the matrix of feature space is U=[u ₁, u ₂..., u _d], training sample is projected on feature space, obtains projection matrix:

Q＝U ^TA (7)

Be the eigenface of sample.

5. the face recognition algorithms of characteristic weighing according to claim 1, is characterized in that: use AHP algorithm to be weighted according to the importance of each component, specifically comprise:

1., first problem stratification, construct a stratified structural model;

2. Judgement Matricies, the relation between hierarchical structure reflection factor, but each criterion proportion in target is weighed is different;

3. Mode of Level Simple Sequence and consistency check, determines the associated element importance of this level, and the weighted value between their order.

6. the face recognition algorithms of characteristic weighing according to claim 1, it is characterized in that: use support vector machine (SVM) to carry out Classification and Identification, specifically comprise: select suitable support vector machine parameter, extract face characteristic data label, the face characteristic data extracted are carried out training and obtains training set, again test sample book is supplied to support vector machine, provides recognition result by the supporting vector machine model trained.